Beyond 'Aha!': Toward Systematic Meta-Abilities Alignment in Large Reasoning Models

TL;DR

This paper introduces a method to explicitly align large reasoning models with core meta-abilities like deduction, induction, and abduction, improving their reasoning performance and reliability across various benchmarks.

Contribution

It proposes a three-stage pipeline for meta-ability alignment using self-verifiable tasks, enhancing reasoning capabilities beyond prompt-based methods.

Findings

Performance improved by over 10% on instruction-tuned baselines.

Domain-specific RL further boosts reasoning performance.

Explicit meta-ability alignment provides a scalable, dependable foundation for reasoning.

Abstract

Large reasoning models (LRMs) already possess a latent capacity for long chain-of-thought reasoning. Prior work has shown that outcome-based reinforcement learning (RL) can incidentally elicit advanced reasoning behaviors such as self-correction, backtracking, and verification phenomena often referred to as the model's "aha moment". However, the timing and consistency of these emergent behaviors remain unpredictable and uncontrollable, limiting the scalability and reliability of LRMs' reasoning capabilities. To address these limitations, we move beyond reliance on prompts and coincidental "aha moments". Instead, we explicitly align models with three meta-abilities: deduction, induction, and abduction, using automatically generated, self-verifiable tasks. Our three stage-pipeline individual alignment, parameter-space merging, and domain-specific reinforcement learning, boosting performance by over 10\% relative to instruction-tuned baselines. Furthermore, domain-specific RL from the aligned checkpoint yields an additional gain in performance ceiling for both 7B and 32B models across math, coding, and science benchmarks, demonstrating that explicit meta-ability alignment offers a scalable and dependable foundation for reasoning. Code is available at: https://github.com/zhiyuanhubj/Meta-Ability-Alignment